The Bühlmann-Straub Hierarchical Model

Why a Hierarchical Model Is Needed

A risk’s observed loss cost reflects two sources of variation. First, the risk has an underlying expected cost that may differ from other risks. Second, actual experience fluctuates around that expectation. The Bühlmann-Straub model represents both sources explicitly, while allowing observations to have different exposure volumes.

The hierarchy is important: a risk has a latent conditional mean, and those latent means vary across a collective of risks. Conditional on its own mean, each risk produces observations with process variation. Exposure determines the scale of that process variation. The model’s assumptions describe these layers; they do not imply that every risk has the same expected cost.

Definition: For risk $i$, let $\Theta_i$ represent its latent risk characteristic. Let $Y_{ij}$ be the observed loss cost per exposure unit in period $j$, and let $w_{ij}>0$ be the corresponding exposure. The risk’s conditional risk mean is $m(\Theta_i)=E[Y_{ij}\mid\Theta_i]$. It is the expected cost for that risk, conditional on its latent characteristic, not the realized observation $Y_{ij}$.

The Bühlmann-Straub Assumptions

The model begins by specifying the conditional distribution of observations for a given risk. In its classical form, observations are conditionally independent given $\Theta_i$, and their first two conditional moments are

E[Y_{ij}\mid\Theta_i]=m(\Theta_i),\qquad \operatorname{Var}(Y_{ij}\mid\Theta_i)=\frac{\phi}{w_{ij}},

where $\phi>0$ is the common process-variance parameter. Thus, for the same risk characteristic, the expected cost per exposure unit does not change merely because the period has more exposure. Its conditional variance does change: multiplying exposure by four divides the variance of the per-unit observation by four. Equivalently, if total loss in the period is $w_{ij}Y_{ij}$, its conditional variance is $w_{ij}\phi$.

The collective assumption describes variation in the conditional means. For risks drawn from the collective,

E[m(\Theta_i)]=\mu,\qquad \operatorname{Var}(m(\Theta_i))=a.

Here $\mu$ is the collective mean of the latent risk means, and $a$ is their variance. The quantity $a$ measures between-risk heterogeneity; it is not the variance of an individual observation. The parameter $\phi$ measures process variation conditional on risk type. The classical model assumes $\phi$ is common across risk characteristics. A more general version allows conditional process variance $\phi(\Theta_i)/w_{ij}$; then the expected process variance per unit exposure is $E[\phi(\Theta_i)]$, but the common-$\phi$ specification no longer holds risk by risk.

The model also assumes the exposures are positive, known, and measured on a consistent basis. Conditional independence means that, after conditioning on a risk’s latent characteristic, one period’s process fluctuation does not predict another’s. It does not say that observations from the same risk are unconditionally independent: they share the same latent mean.

Model checklist: Identify the risk-level latent characteristic; state the conditional mean; state how conditional variance scales with exposure; specify conditional independence; and describe the collective mean and variance of the conditional risk means. In the classical Bühlmann-Straub model, the process parameter $\phi$ is common across risks.

How the Two Sources of Variation Combine

The law of total variance shows why process variance and risk heterogeneity must be kept distinct. For one observation with exposure $w$, the conditional mean is $m(\Theta)$ and the conditional variance is $\phi/w$. Therefore,

\begin{aligned} \operatorname{Var}(Y) &=E[\operatorname{Var}(Y\mid\Theta)] +\operatorname{Var}(E[Y\mid\Theta])\\ &=E\left[\frac{\phi}{w}\right] +\operatorname{Var}(m(\Theta))\\ &=\frac{\phi}{w}+a. \end{aligned}

This calculation assumes a fixed exposure $w$ and the classical common-$\phi$ model. It explains the two components: $\phi/w$ is within-risk process variance, while $a$ is variation in expected costs across risks. Increasing exposure reduces the first component, but does not remove the second.

For two distinct periods $j\ne k$ of the same risk, conditional independence also gives a useful interpretation. Their conditional covariance is zero, but their unconditional covariance is

\begin{aligned} \operatorname{Cov}(Y_{ij},Y_{ik}) &=E[\operatorname{Cov}(Y_{ij},Y_{ik}\mid\Theta_i)]\\ &\quad+\operatorname{Cov}(E[Y_{ij}\mid\Theta_i], E[Y_{ik}\mid\Theta_i])\\ &=0+\operatorname{Var}(m(\Theta_i))=a. \end{aligned}

The shared latent risk mean creates the dependence across periods. This result requires distinct conditionally independent observations and a common risk mean across the periods. It would not apply unchanged if the underlying risk level changed over time or if residual dependence remained after conditioning.

Worked Example: Specifying a Two-Type Collective

A hypothetical insurer studies a portfolio with two equally likely latent risk types. Conditional on type A, the annual loss cost per exposure unit is normally distributed with mean $80; conditional on type B, it is normally distributed with mean $120. In either type, an observation at exposure $w$ has conditional variance $400/w$. Find the collective mean and variance of conditional risk means, then find the marginal variance of an observation at exposure $w=4$.

Method and assumptions. Use the two-level model: calculate the mean and variance of the type-specific conditional means, then add the conditional process variance to the between-risk variance. The conditional distribution is $Y\mid A\sim N(80,400/w)$ and $Y\mid B\sim N(120,400/w)$, where the second normal parameter is the variance. Interest and mortality assumptions do not apply to this loss-cost model.

The conditional risk means are 80 and 120, each with probability $0.5$. Thus,

\mu=0.5(80)+0.5(120)=40+60=100.

The between-risk variance is

$$ a=0.5(80-100)^2+0.5(120-100)^2 =0.5(400)+0.5(400)=400. $$

At exposure $w=4$, the conditional process variance is $400/4=100$, for either type. Hence the marginal variance is

\operatorname{Var}(Y)=a+\frac{\phi}{w}=400+100=500.

Check. The marginal mean is 100, the midpoint of the equally likely type means. The total variance includes both the 400 units of type variation and the 100 units of process variance. The common mistake is to report only $400/4=100$, which omits differences in expected cost across risks.

Exposure Scaling and Model Diagnostics

Exposure is not simply another weight attached to the observation. It changes the conditional variance under a specific model assumption. If a per-unit observation is based on four times as much exposure, the model says its process variance is one-quarter as large, provided the risk characteristic and the exposure basis are otherwise comparable. A mismatch between the exposure definition and the quantity being modeled can therefore invalidate the variance specification.

The assumptions can be assessed separately. If observations at equal exposure show systematically different conditional variances across risk types, a common $\phi$ may be inappropriate. If residual outcomes remain correlated across periods after conditioning on risk characteristics, conditional independence may fail. If a risk’s expected cost drifts over time, a single fixed $m(\Theta_i)$ may not describe all periods. These are model limitations, not reasons to reinterpret $a$ as process variance.

A model can preserve the inverse-exposure relationship while allowing different process variances by risk type. In that case, write $\operatorname{Var}(Y_{ij}\mid\Theta_i)=\phi(\Theta_i)/w_{ij}$. This is a broader hierarchical specification. It should not be silently treated as the classical common-$\phi$ model, because the conditional process variance now depends on the latent risk characteristic.

Worked Example: A Risk-Dependent Process Variance

A hypothetical equipment portfolio has two equally likely risk types. Type L has conditional mean loss cost $50 and process-variance parameter 100; type H has conditional mean $100 and process-variance parameter 900. At exposure $w$, each type’s conditional observation is normal with its stated mean and variance equal to its process parameter divided by $w$. Calculate the collective mean, between-risk variance, average process parameter, and marginal variance at $w=2$. Determine whether the common-$\phi$ assumption holds.

Method and assumptions. Use the generalized exposure-scaled model with $\phi(\Theta)=100$ for L and 900 for H. The conditional distributions are $Y\mid L\sim N(50,100/w)$ and $Y\mid H\sim N(100,900/w)$. Interest and mortality assumptions do not apply.

The collective mean is

\mu=0.5(50)+0.5(100)=25+50=75.

The conditional risk means differ from 75 by $-25$ and $25$, so

$$ a=0.5(-25)^2+0.5(25)^2 =0.5(625)+0.5(625)=625. $$

The average process-variance parameter is

E[\phi(\Theta)]=0.5(100)+0.5(900)=50+450=500.

At $w=2$, the marginal process variance is $500/2=250$. Applying total variance gives

\operatorname{Var}(Y)=625+250=875.

Check. Directly, the two conditional variances at $w=2$ are 50 and 450. Their average is 250, and the between-type variance is 625, giving 875. The common-$\phi$ assumption does not hold because the conditional process parameters are 100 and 900, not one shared value. Using $\phi=500$ would reproduce the average process contribution but would not make the classical conditional variance assumption true for each type.

Worked Example: Checking Conditional Independence

For a hypothetical risk with a fixed latent mean $m(\Theta)$, suppose two period observations have exposure 1 and conditional process variance 300 each. Their conditional covariance, given $\Theta$, is zero. Across risks, the conditional means have variance 200. Find the covariance between the two observations across the collective, and explain what it represents.

Method and assumptions. Apply the law of total covariance, assuming the two observations are conditionally independent and share the same latent risk mean. The stated conditional process variances are 300 in each period; the conditional distribution need not be specified beyond its moments. Interest and mortality assumptions do not apply.

The law of total covariance is

\operatorname{Cov}(Y_1,Y_2) =E[\operatorname{Cov}(Y_1,Y_2\mid\Theta)] +\operatorname{Cov}(E[Y_1\mid\Theta],E[Y_2\mid\Theta]).

The first term is zero by conditional independence. Both conditional expectations equal $m(\Theta)$, so the second term is its variance, 200. Therefore,

\operatorname{Cov}(Y_1,Y_2)=0+200=200.

Check. The covariance is not 300: that is a conditional variance within one period, not the cross-period covariance. The positive unconditional covariance arises because risks with higher latent means tend to have higher observations in both periods.

Common Exam Traps

Confusing the two variance components. $a$ describes variation among conditional risk means; $\phi/w$ describes conditional process variation. They answer different questions.
Scaling the wrong quantity. The model states that the variance of the per-unit observation is $\phi/w$, not $\phi w$. For total loss $wY$, the corresponding variance is $w\phi$.
Assuming conditional independence means marginal independence. Distinct observations from the same risk generally share $m(\Theta)$, producing covariance $a$ under the classical assumptions.
Replacing risk-specific process parameters with their average without qualification. An average process parameter can describe an average variance contribution, but it does not establish a common conditional variance across risks.
Treating exposure as a cure for heterogeneity. Greater exposure reduces process variance; it does not reduce the collective variance $a$ of latent risk means.

Key takeaway: Bühlmann-Straub separates a risk’s latent expected cost from the process noise around it. Exposure scales the conditional process variance inversely, while the collective distribution of conditional means describes how risks differ. Check each assumption independently before applying results built on the model.

Check Your Understanding

For each problem, identify the relevant model layer and keep conditional quantities distinct from collective quantities.

A collective has equally likely risk types with conditional means $40 and $80. Find $\mu$ and $a$. (Answer: $\mu=60$, $a=400$.)
Under the classical model, $\phi=720$, exposure is $w=6$, and $a=250$. Find the conditional process variance and marginal variance of one observation. (Answer: process variance $120$; marginal variance $370$.)
Two distinct observations for one risk are conditionally independent and share the same latent mean, whose collective variance is 150. What is their unconditional covariance? (Answer: $150$.)
Two equally likely types have means 30 and 90, and process parameters 200 and 600. At exposure $w=4$, find $\mu$, $a$, average process variance, and marginal variance; state whether the common-$\phi$ assumption holds. (Answer: $\mu=60$, $a=900$, average process variance $100$, marginal variance $1{,}000$; no, the process parameter differs by type.)
A per-unit loss-cost observation has conditional variance 240 at exposure 1 under the classical model. What variance does the model assign at exposure 5, and what is the variance of total loss at exposure 5? (Answer: per-unit variance $48$; total-loss variance $1{,}200$.)

What You'll Learn